Candidate: Biniyam Deressa
Date: January 30, 2026
Time: 2:00pm
Location: Online—contact the candidate for more details.
Supervisor: Dr. Anwar Hasan
All are welcome!
Abstract:
Federated learning enables collaborative model training without centralizing data, but it relies on the unverified assumption that clients faithfully perform local computation. In practice, participants may free ride or submit fabricated updates, undermining fairness and model integrity, especially in incentive-driven or high-stakes deployments. Existing defenses face a fundamental tension: cryptographic proofs are too expensive for realistic training workloads, while statistical or heuristic methods provide no formal guarantees.
This talk presents cryptographically verifiable training, focusing on Retroactive Intermediate Value Verification (RIV), a non-interactive proof of training protocol based on polynomial commitments and numerical error analysis. RIV enables verification of forward and backward passes without quantization or circuit-level encodings and prevents selective computation by committing before challenges are revealed. The talk emphasizes intuition, system design, and practical implications, and briefly connects these ideas to recent work on succinct verification of core cryptographic and machine learning computations.